DigiFarm's solution will enable highly accurate and unbiased classification of agricultural land using up-to-date super-high resolution Satellite data (1m per pixel) in order to automatically detect field boundaries and seeded acres.
The solution will address the "Property Registration" section of the challenge. Furthermore, the solution will enable smallholder farmers to easier access financing, agronomic advisory and build their credit profile. Furthermore, with nearly 80% of households in Tanzania engaging in agriculture and at least one third gaining more than half of their income from agricultural activities, access to finance for small-scale producers is a major catalyst to broad based economic growth.
For a long time and especially in traditional forms of financing, one of the key limiting factors for access to loans for smallholder farmers has been lack of collateral. Looking at land ownership registration for instance, data collected by Tanzania’s bureau of statistics in 2018 shows that out of 8.7 million farms surveyed only 18% were registered.
The current problem is the lack of historical and in-season data to assess credit risk on smallholder farmers, this due to a lack of infrastructure consisting of agricultural land classification, crop classification, long-term productivity assessment (20-30+ years) on the individual farm-land and field boundaries.
Currently, field boundaries are manually created by field agents walking the corners of a physical agricultural field and geo-tagging those boundaries, this is time-consuming, expensive and often inaccurate. In order to provide a reliable and affordable solution the only way is to use automate this through the use of deep learning object detection and high-resolution Satellite data.
DigiFarm's solution will enable and empower smallholder farmers through open access to:
- Detecting in-season field boundaries and seeded acres - will help farmers assess and plan their growing season accurately through accessing the crop-inputs required (seeds, fertiliser, crop protection).
- Long-term productivity assessment - DigiFarm will provide 35 years of long-term productivity data (NDVI) for each individual field, including slope (growth of plant biomass over that period) which can be used to access financing and build a credit rating.
- In-season detection of management zones - DigiFarm will provide in-field assessment of the low, medium and high productivity zones (EVI/biomass) in order for the farmer to assess the problem areas and ability to apply variable fertiliser/fungicide application.
The main risks associated with this project is:
Type of risk: Technological
Description of the risk: Inability for deep neural network models to not accurately detect and predict field boundaries and seeded acres at a high accuracy in smallholder markets such as Tanzania where field boundaries are typically small and dynamic, due to extensive agricultural practices differences and variability, topography, crops, agricultural practice and weather (cloud-cover) restricting the model to interpret the imagery-data accurately, and accuracy resulting in lower than 0.96 (Intersection over Union) for field boundary detection.
Effect of the risk? Effect of risk includes that the field boundary detection model does not reach high enough accuracy in order to be commercially attractive to clients, i.e. not high enough accuracy compared to Cadastral map data or manual delineation.
Mitigation methods? Solution; develop regional models which provide the highest level of accuracy for key markets, to collect a substantial amount of training data (manual delineation) for the targeted areas. Additionally, DigiFarm super-resolves Sentinel-2 from 10m to 1m per pixel resolution which will be critical in order to properly delineate the field boundaries.
Type of risk: Cost
Description of risk: computational complexity and costs of data processing too high for planned price-point and practical market applications, limiting uptake and market segment.
Effect of the risk: if data processing costs are too high which will reduce our profit margins, this will risk the long-term sustainability of DigiFarm’s business model and operational capacity, as DigiFarm has identified the price-point with extensive market research to determine the (lowest possible price), currently 25% lower than competitors for 10x higher resolution and accuracy.
Mitigation methods: Data engineering and algorithmic design is continuously focusing on computational efficiency and scalability. DigiFarm will set up its own HPC infrastructure with local GPU-instances in order to reduce costs of data processing in cloud providers ecosystems where costs are ~$3 per hour per GPU, compared to $0.30 per hour. This has only been made possible due to: (a) recent advances in hardware (Graphic Processing Units) by industry (lead by NVIDIA) where performance of GPUs have tripled since 2015, (200 TFLOPS to 1500) combined with the affordability of the units (approx. 300% reduction in performance/$ ratio).